SVD did not converge: #143 (cont, supposedly fixed by #151) about autode HOT 5 OPEN

Hi,

Any updates about this?

Thanks,

E77

from autode.

Hope you don't mind but I edited the issue to make the error/code easier to read

Thanks for raising –this looks like an optimiser bug to me, probably present since v1.3.0 – I'll take a look. I'm hoping @shoubhikraj might have fixed it already in #262

from autode.

I cant reproduce this error but there's definitely something bad going on – it shouldn't take 68 iterations to converge N2O3 & fail to run the internals->cartesians transform 😞

@shoubhikraj would you mind trying ade.Molecule(smiles='[O-][N+](=O)N=O').optimise(method=ade.methods.NWChem()) with the trim optimiser when you have the chance? thanks!

from autode.

@t-young31 @eneas77

I do not have access to NWChem right now, but I can reproduce the error even with xTB and the new optimiser:

>>> import autode as ade
>>> mol = ade.Molecule(smiles='[O-][N+](=O)N=O')
>>> opt = ade.opt.optimisers.HybridTRMOptimiser(maxiter=100,gtol=1.e-3,etol=1.e-4)
>>> opt.run(mol, method=ade.methods.XTB())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "G:\autodE\autode\opt\optimisers\base.py", line 135, in run
    self._step()  # Update self._coords
  File "G:\autodE\autode\opt\optimisers\trm.py", line 174, in _step
    self._coords = self._coords + step  # finally, take the step!
  File "G:\autodE\autode\opt\coordinates\base.py", line 264, in __add__
    new_coords.iadd(other)
  File "G:\autodE\autode\opt\coordinates\dic.py", line 243, in iadd
    raise RuntimeError(
RuntimeError: Iterative back-transform did not converge
>>> mol.coordinates
Coordinates([[-1.38900795 -0.83276619  0.04883061]
 [-0.58297449  0.02862588 -0.01147048]
 [-0.67159798  1.20019831 -0.15331935]
 [ 0.90290826 -0.58392015  0.17367619]
 [ 1.74067215  0.18786215 -0.05771697]] Å)

The optimiser throws the error much earlier (~16 iterations) when the dic-> cartesian transform fails. I suspect the problem is coming from the nearly 180 degree dihedral angle. When I restart the optimisation (which rebuilds the internal coordinates) it works (I assume because the dihedrals are now excluded from the primitive set):

>>> opt = ade.opt.optimisers.HybridTRMOptimiser(maxiter=100,gtol=1.e-3,etol=1.e-4)
>>> opt.run(mol, method=ade.methods.XTB())
>>> mol.coordinates
Coordinates([[-1.44320967e+00 -8.13930256e-01  8.45857398e-02]
 [-5.93660808e-01 -5.68579976e-04 -1.19083114e-02]
 [-5.96507906e-01  1.17034244e+00 -1.76402679e-01]
 [ 9.38350317e-01 -6.14585625e-01  1.07541973e-01]
 [ 1.69502807e+00  2.58742025e-01 -3.81672248e-03]] Å)

If you look at the final optimised structure, the two O-N-N-O dihedral angles are 179.773 and -0.264 degrees:

One way to solve this could be to modify the optimisers so that whenever the back-tranform fails to converge, it will rebuild internal coordinates or perhaps switch to Cartesians.

from autode.

Hi,

Did you try to run this example when you closed the enhancement on #151 ?
I suppose the answer, as there is no one, is you didn't run it.

Perhaps you should include that basic QC: every case ending in bug/enhancement should be tested once the fork job is done. :-)

Just my two shilings on this matter.

Thanks a lot.

E77

from autode.